The company claims the model demonstrates performance comparable to GPT-5.2-Thinking, Claude-Opus-4.5, and Gemini 3 Pro.
You can’t cheaply recompute without re-running the whole model – so KV cache starts piling up Feature Large language model ...